The state of agentic coding is one of my favorite new podcasts of late. The name sounds just like the same slop as everything else, but Ben and Armin are top tier hosts that actually do this stuff for work and on the side.
slowing down and speeding up #
Which brings me to the first interesting topic they talked about. Both talked about going slow at work. Ben still reads everything and deeply understands the inside of what he is working on. Home is a different story. At home both are cranking on new side projects, vibing them out quick and fast. The projects work for them and that’s all they need. They dont need to support hundreds of users. All making it much easier to just let the clanker kick new things out fast.
open weight #
An interesting side topic on open weight models. Armin opened with them being too expensive. The big labs have such good subsidies right now that it’s hard to justify not using them. He did have an interesting use case though. He has a side project using cheap Chinese carplay devices. When trying to get openai to work on them it tends to refuse to work for security reasons, but if he starts the session with kimi, openai has no problems continuing.
This is next level of whispering just the right tokens to the model to get it to magically do things for you that it otherwise wouldn’t.
He also commented this tends to work with distillation refusals too. He said he will get refusals for false distillation attempts when working on pi.
new tools?? #
Armin admits that he tries all of the new ai tools as they come out, maybe not all, but a lot. He says he has adopted a whopping zero of them and his AI tools have not changed in the last year. They just don’t seem to add anything for him. Agents with a good harness and some skills is about all he needs.